[I accidentally sent earlier reply by mail only, sorry for duplicates:] Francis Tyers <fty...@prompsit.com> wrote: > El 2018-08-04 19:44, Xavi Ivars escribió: >> This is great! >> >> However, is a bit counter-intuitive that lower weight wins. In other >> cases (like lexical selection rules), higher weights win.
(I actually always thought it was lower weight for lexical selection as well) >> Would it be possible to do something so weights work consistently? >> > > This is a good question. And one that probably should be addressed > by Flammie or Abinash. > > Guys, what do you think ? Is it possible to make the weights > "higher=better", It might be doable engineering-wise, I'm not totally sure if it works in finite-state algebra though. One thing to remember is that these weights underlyingly and in practice also are defined for basically every character in the string and combined, even in the system it's for all morphs not just roots, and the way they are added together is with the addition +, it's very easy to make intuitive systems that way but with low weight system it would be way more complicated, to tell what does it mean to take lexeme of weight 1, 2 or 3 and suffix morphs of weights 1 or 2 or 3. With the weight is bad interpretation we get automatically the "more morphs is worse" that is quite ok default linguistics-wise. There is a -signed weight system in e.g. openfst that was added exactly for this intuition of weights, so it operates on R-, min, - instead of R+, max, +, maybe it could be an ok solution? I think the background of this weight logic is is the penalty weight interpretation is very dominant in CS data structures and algorithms and the way it's given in the basic courses alongside e.g. travelling salesman, most of the CS students won't even think weight to be a good thing in anything, after all we don't want to be heavy or have heavy luggage or travel in heavy terrains etc. And consequently of course all the systems like HFST openfst etc. are designed with this weight interpretation and it's unlikely to change. > this would also make it easier to use probabilities I suppose. Nah, the tropical weights very pretty much designed with -log(prob) in mind already. -- Flammie, computer scientist bachelor + linguist master = computational linguist doctor, free software Finnish localiser, and more! <http://www.iki.fi/flammie/> ------------------------------------------------------------------------------ Check out the vibrant tech community on one of the world's most engaging tech sites, Slashdot.org! http://sdm.link/slashdot _______________________________________________ Apertium-stuff mailing list Apertium-stuff@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/apertium-stuff
